MCP Weather & Transit Advisor

A small Model Context Protocol server that lets Claude check real-time weather and plan real public-transport trips, then recommend a route that suits the weather.

Tools

Related MCP server: open-meteo

Data sources

Transit coverage depends on which operators publish open data. Singapore, most of Europe and much of North America are covered; some regions are not. Times marked scheduled are timetabled, live include real-time updates.

Optional: live Singapore bus arrivals

1. Request a free AccountKey at https://datamall.lta.gov.sg/content/datamall/en/request-for-api.html.

2. Provide it as the LTA_ACCOUNT_KEY environment variable (see the Claude Desktop config below).

Without a key, get_stop_departures falls back to Transitous timetables for Singapore stop codes.

Caching

API answers are cached in a local SQLite file (~/.mcp-weather-advisor/cache.sqlite3) so repeated questions are answered instantly and the free APIs aren't hammered. Cached answers end with "(Served from cache, fetched N min ago.)".

• Change the TTLs at the top of mcp_weather_advisor/server.py (WEATHER_TTL, TRANSIT_TTL, ...).

• WEATHER_ADVISOR_CACHE=/path/to/file.sqlite3 stores the cache elsewhere; WEATHER_ADVISOR_CACHE=off disables it.

• API keys are never written to the cache. If the cache file can't be opened, the server simply calls the APIs directly.

• To clear the cache, delete ~/.mcp-weather-advisor/cache.sqlite3.

Installation

1. Install Python 3.10 or newer

• macOS: download from https://www.python.org/downloads/ (or brew install python).

• Windows: download from the same page and tick "Add python.exe to PATH" in the installer.

• Check it worked: python3 --version (on Windows: python --version).

2. Clone the project and create a virtual environment

A virtual environment (.venv) is a private Python install for this project. Create it inside the project folder: the Claude Desktop config below points at it.

macOS / Linux

git clone https://github.com/viaanshetty3-debug/mcp-weather-advisor.git
cd mcp-weather-advisor
python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Windows (PowerShell)

git clone https://github.com/viaanshetty3-debug/mcp-weather-advisor.git
cd mcp-weather-advisor
python -m venv .venv
.venv\Scripts\Activate.ps1
pip install -e .

pip install -e . reads pyproject.toml and:

• installs the dependencies (mcp, httpx);

• installs the server as an editable package, so edits to the code take effect without reinstalling;

• adds a mcp-weather-advisor command that starts the server.

Installing without a virtual environment: pip install -e . also works system-wide, but some Pythons (Homebrew's, and many Linux distributions') refuse with an externally-managed-environment error. There, use the virtual environment above, or pipx: pipx install -e ..

Intel Mac note: cryptography (a dependency of mcp) stopped publishing Intel Mac wheels after version 48. pyproject.toml therefore keeps it below 49 on Intel Macs only, so the install doesn't try to compile it from source. Other platforms get the latest version.

3. Check it runs

python server.py

The server starts and waits silently for Claude to talk to it over stdio. That's expected; press Ctrl+C to stop. mcp-weather-advisor and python -m mcp_weather_advisor start the same server.

To try the tools interactively in the MCP Inspector (needs Node.js and pip install -e ".[dev]"):

mcp dev mcp_weather_advisor/server.py

Connect to Claude Desktop

Claude Desktop starts the server itself, using two absolute paths from its config file:

Step 1: Get your exact block (recommended)

With the virtual environment activated, run this from the project folder (macOS/Linux). It prints the block with your real paths filled in:

python -c 'import json, os, sys; print(json.dumps({"mcpServers": {"weather-transit-advisor": {"command": sys.executable, "args": [os.path.abspath("server.py")]}}}, indent=2))'

Or write it by hand. Find the project folder's absolute path with pwd (macOS/Linux) or (Get-Location).Path (Windows), then fill in one of the templates below.

Step 2: The configuration block

macOS / Linux (replace /Users/alex/mcp-weather-advisor with your project folder):

{
  "mcpServers": {
    "weather-transit-advisor": {
      "command": "/Users/alex/mcp-weather-advisor/.venv/bin/python",
      "args": [
        "/Users/alex/mcp-weather-advisor/server.py"
      ]
    }
  }
}

Windows (replace C:\\Users\\alex\\mcp-weather-advisor, keeping the double backslashes that JSON requires):

{
  "mcpServers": {
    "weather-transit-advisor": {
      "command": "C:\\Users\\alex\\mcp-weather-advisor\\.venv\\Scripts\\python.exe",
      "args": [
        "C:\\Users\\alex\\mcp-weather-advisor\\server.py"
      ]
    }
  }
}

Optional: live Singapore bus arrivals. If you have an LTA DataMall key, add an env block after args. Remember the comma after the closing ]:

{
  "mcpServers": {
    "weather-transit-advisor": {
      "command": "/Users/alex/mcp-weather-advisor/.venv/bin/python",
      "args": [
        "/Users/alex/mcp-weather-advisor/server.py"
      ],
      "env": {
        "LTA_ACCOUNT_KEY": "paste-your-real-key-here"
      }
    }
  }
}

Leave env out entirely if you don't have a key. A placeholder value causes 401 errors for Singapore stop codes.

Step 3: Paste it into Claude Desktop

1. In Claude Desktop, open Settings → Developer → Edit Config. This opens claude_desktop_config.json:

   • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

   • Windows: %APPDATA%\Claude\claude_desktop_config.json

2. Paste the block in:

   • If the file is empty or {}: replace its contents with the block.

   • If it already has an "mcpServers" section: don't paste a second one. Copy only the "weather-transit-advisor": { ... } entry into the existing "mcpServers" object, with a comma between it and the server before it:

     {
       "mcpServers": {
         "some-other-server": { "command": "..." },
         "weather-transit-advisor": {
           "command": "/Users/alex/mcp-weather-advisor/.venv/bin/python",
           "args": ["/Users/alex/mcp-weather-advisor/server.py"]
         }
       }
     }

3. Save, then fully quit Claude Desktop (macOS: ⌘Q; Windows: right-click the tray icon → Quit) and reopen it. Just closing the window isn't enough.

4. The tools appear under the tools (🔨) menu in a new chat. Try:

   • "What's the weather in Lisbon, and should I cycle to work?"

   • "How do I get from Orchard MRT to Changi Airport right now?"

   • "When's the next bus at stop 83139?"

Common mistakes

Claude Desktop's own log for this server is ~/Library/Logs/Claude/mcp-server-weather-transit-advisor.log on macOS (Windows: %APPDATA%\Claude\logs\). It shows startup errors such as the ones above. Problems inside the tools go to this project's own log; see Logging.

Alternative: instead of python + server.py, command can be the installed launcher alone (/Users/alex/mcp-weather-advisor/.venv/bin/mcp-weather-advisor, Windows ...\\.venv\\Scripts\\mcp-weather-advisor.exe) with no args. Both start the same server.

Project layout

mcp-weather-advisor/
├── server.py                 ← the file Claude Desktop runs (a small launcher)
├── mcp_weather_advisor/      ← the server package
│   ├── __init__.py           version number
│   ├── __main__.py           enables `python -m mcp_weather_advisor`
│   └── server.py             the server: settings, logging, cache, helpers, the four tools
├── tests/                    pytest suite (fake APIs; no internet needed)
│   ├── conftest.py           fixtures: starts the server over stdio
│   ├── fake_api.py           local stand-in for Open-Meteo / Transitous / LTA
│   ├── sample_data.py        realistic API responses for the tests
│   └── test_*.py             the tests (see Testing)
├── pyproject.toml            packaging: dependencies, `mcp-weather-advisor` command, pytest settings
├── requirements.txt          shortcut for `pip install -e .`
├── requirements-dev.txt      shortcut for `pip install -e ".[dev]"`
├── README.md
├── LICENSE                   MIT
└── .venv/                    your virtual environment (created by you, not in git)

mcp_weather_advisor/server.py is organised in labelled sections: settings, logging, cache, helpers, the four tools (decorated with @mcp.tool()) and the entry point. Each tool's docstring is what Claude reads to decide when to use it.

Where files are stored

The cache (cache.sqlite3) and log (mcp_server.log) live in ~/.mcp-weather-advisor/, created on first use. They're kept outside the code folder because an installed package's folder may be read-only. Set WEATHER_ADVISOR_HOME to use another folder, or move each file individually with WEATHER_ADVISOR_CACHE and WEATHER_ADVISOR_LOG.

Logging

Problems are written to ~/.mcp-weather-advisor/mcp_server.log as JSON Lines, one JSON object per line:

{"ts": "2026-11-10T03:29:38.790Z", "level": "ERROR", "event": "api_timeout", "tool": "get_current_weather", "call_id": "27efa384", "url": "https://geocoding-api.open-meteo.com/v1/search", "params": {"name": "tokyo", "count": 1, "language": "en"}, "timeout_s": 20.0, "elapsed_ms": 20004, "error": "ReadTimeout"}

Every record has ts (UTC), level and event. Records made during a tool call also carry tool and a call_id, so you can group everything from one call.

• Nothing is ever printed to stdout. stdout carries the MCP protocol, and the test suite checks it stays clean while every failure type is logged.

• No failure crashes a tool. Every one becomes a readable message for Claude, such as "Could not get weather for 'Tokyo': the service did not respond within 20 seconds."

• The log rotates at about 1 MB, keeping mcp_server.log.1 to .3.

• API keys are never logged. Request headers are not recorded, and query parameters with names like key, token or secret are redacted.

• WEATHER_ADVISOR_LOG=/path/to/file.log moves the log; WEATHER_ADVISOR_LOG=off disables it. If the file can't be opened, log lines go to stderr instead, which Claude Desktop saves in its own MCP logs.

• WEATHER_ADVISOR_TIMEOUT sets the network timeout in seconds (default 20).

Quick look at recent errors (needs jq):

jq -c 'select(.level != "INFO") | {ts, event, tool, error, status}' ~/.mcp-weather-advisor/mcp_server.log

Testing

pip install -e ".[dev]"
python -m pytest

The suite (~255 tests, about 90 seconds) needs no internet and no API keys:

• Over stdio, like Claude Desktop: almost every test starts the installed server as a subprocess and calls its tools with the official MCP client.

• Against a local fake API: each test scripts the fake's replies, via the environment variables OPEN_METEO_GEOCODING_URL, OPEN_METEO_FORECAST_URL, TRANSITOUS_URL and LTA_BUS_ARRIVAL_URL.

Continuous integration

.github/workflows/tests.yml runs the whole suite on GitHub Actions for every push and pull request (and on demand from the Actions tab), on Python 3.11 and 3.12 on Ubuntu. Each run:

1. installs the package with pip install -e ".[dev]", exactly as a developer would;

2. runs the SQLite cache tests (tests/test_cache.py) as their own step;

3. runs everything else: the MCP stdio tool tests, API-failure handling, logging, packaging and unit tests. This step runs even if the cache step failed, so all problems show up in one run.

If anything fails, the run keeps a downloadable artifact with JUnit reports plus each test's server stderr and mcp_server.log. No secrets are needed: the tests use a local fake API.

Troubleshooting

• Server doesn't appear in Claude: see Common mistakes.

• ModuleNotFoundError: run pip install -e . with the .venv activated, and make sure command is that .venv's Python (see Common mistakes).

• mcp-weather-advisor: command not found: the environment's bin (or Scripts) folder isn't on your PATH. Activate the virtual environment, or use the full path.

• "No location found": try a larger nearby city or add the country.

• "No public-transport route found": the region may not publish open transit data on Transitous.

• LTA errors (401): the LTA_ACCOUNT_KEY is missing or wrong.

License

MIT. Weather data by Open-Meteo.com (CC BY 4.0). Transit routing by Transitous using open GTFS data from each operator. Singapore bus arrivals from LTA DataMall under the Singapore Open Data Licence.